Can deep learning separate two electrochemical signals that overlap and change with time?

The electrochemical problem

Fast-scan cyclic voltammetry can follow neurotransmitter release with subsecond time resolution, but chemically similar species can generate overlapping oxidation and reduction features. Dopamine and octopamine are a particularly difficult pair in Drosophila because octopamine has a secondary oxidation process near the dopamine oxidation potential, and the voltammetric shape of octopamine changes with time.

That time dependence matters. Methods built around a fixed basis spectrum can struggle when the underlying electrochemical signature itself evolves. The paper therefore asks whether a model that learns the two-dimensional time–voltage–current pattern can separate the signals without assuming that every voltammogram has a fixed shape.

What the researchers did

The authors compared deep-learning architectures and selected a modified U-Net regression model. A distinct primary oxidation feature of octopamine near 1.1 V was used to predict its secondary oxidation contribution near the region that overlaps dopamine. The predicted octopamine contribution could then be subtracted from a mixed signal to recover the dopamine component.

The strategy was tested first with computational mixtures, where the original dopamine and octopamine signals provide a ground truth, and then with experimentally measured mixtures in the Drosophila brain.

What they found

For computational mixtures, the reported normalized root-mean-square errors were 0.06 for dopamine and 0.08 for octopamine. In experimentally measured 1:1 mixtures, the predicted dopamine and octopamine concentrations showed strong agreement, with a Pearson correlation coefficient of 0.93 and a concordance correlation coefficient of 0.93.

The result is important because the network is not merely assigning a label to a voltammogram. It reconstructs and subtracts a time-varying electrochemical contribution that would otherwise obscure the second analyte.

Why this matters

This is a useful example of AI addressing an electrochemical measurement problem rather than replacing the chemistry. The informative feature is still electrochemical: a distinct octopamine oxidation peak. The neural network learns how that observable feature maps onto the overlapping part of the signal.

The broader opportunity is signal deconvolution in systems where species overlap and their response shapes are not constant in time. That could be relevant well beyond this particular neurotransmitter pair.

One important caution

The model is trained on signals collected under defined electrode, buffer, and biological conditions. The paper itself shows some degradation when testing on electrodes excluded from training. This is a reminder that successful electrochemical signal separation can depend strongly on whether the training data represent the electrodes and environments encountered at inference time.

For the experimentally prepared mixtures there is also no direct ground-truth decomposition of each component, so validation necessarily relies on reference recordings and agreement metrics rather than exact component-wise error.